mixdog 0.9.3 → 0.9.5

package/scripts/bench/round-r18-prewarm.json

@@ -0,0 +1,468 @@
+{
+  "round": "18",
+  "runner": "lead",
+  "opts": {
+    "provider": "openai-oauth",
+    "model": "gpt-5.5",
+    "effort": null,
+    "fast": false,
+    "env": {
+      "MIXDOG_DATA_DIR": "C:\\Users\\tempe\\.mixdog\\bench-data"
+    }
+  },
+  "tasks": 6,
+  "completed": 6,
+  "completion_rate": 100,
+  "sessions": [
+    "sess_24800_1783060135637_1783060135875_396d1f1283f84451dc348604178b23f8",
+    "sess_22404_1783060135650_1783060135889_630e370d8ff15d6c58914f3e6ade30c3",
+    "sess_10516_1783060135671_1783060135907_3c455a35c2d88303ec51be56689a9e3c",
+    "sess_35760_1783060217488_1783060217721_527b673486b2ddb9ace2cd8196d6f95a",
+    "sess_34948_1783060223630_1783060223848_a086c7c3dc32e8d9d25fdba3a4b15cf9",
+    "sess_44084_1783060296498_1783060296721_5059cdd1f7776ce3ddcb030a83e15e87"
+  ],
+  "results": [
+    {
+      "id": "r4-debug-forwarding-gap#1",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 160816,
+      "sessionId": "sess_24800_1783060135637_1783060135875_396d1f1283f84451dc348604178b23f8",
+      "output": "sessionId: sess_24800_1783060135637_1783060135875_396d1f1283f84451dc348604178b23f8\n재영님, 조사 결과 현재 가장 큰 위험은 “이미 존재하는 새 transcript”에 바인딩할 때 EOF로 시작해 초기 출력이 영구 누락될 수 있다는 점입니다.\n\n- **영구 누락 가능**: `bindPersistedTranscriptIfAny()` → `rebindTranscriptContext(... catchUpFromPersisted:true)`가 새 기존 파일에 붙으면 `setContext()`가 `currentSize`를 cursor로 잡아 이미 써진 출력은 스킵합니다. Backlog catch-up 불가. `src/runtime/channels/index.mjs:625`, `src/runtime/channels/lib/output-forwarder.mjs:117`\n\n- **지연 후 복구**: transcript가 아직 없으면 `fs.watch()` 설치가 실패하지만 re-arm poll이 파일 생성 후 `startWatch()` + `forwardNewText()`를 호출합니다. 60초 안이면 backlog catch-up 가능, 이후엔 다음 rebind 없이는 지연/미전송입니다. `src/runtime/channels/index.mjs:469`, `src/runtime/channels/index.mjs:497`, `src/runtime/channels/index.mjs:376`\n\n- **pre-connect 소유권 게이트 지연**: bridge-state는 먼저 쓰이고 backend 연결 전에는 `_isOwner()`가 false라 forwarding이 안 됩니다. cursor는 advance되지 않으며, 연결 후 `forwardNewText()`가 호출돼 backlog를 회수합니다. `src/runtime/channels/index.mjs:1735`, `src/runtime/channels/index.mjs:776`, `src/runtime/channels/index.mjs:809`, `src/runtime/channels/lib/output-forwarder.mjs:351`\n\n- **stat→watch 설치 경계**: `setContext()` 이후 `fs.watch()` 설치 사이에 쓰인 bytes는 `scheduleWatchFlush()`가 즉시 한 번 읽어 회수합니다. Backlog catch-up 가능. `src/runtime/channels/lib/output-forwarder.mjs:112`, `src/runtime/channels/lib/output-forwarder.mjs:665`\n\nVerdict: 기존 파일에 새로 붙는 non-persisted 바인딩만 실제 drop 위험이고, 나머지 활성화 창은 대부분 지연 후 catch-up 구조입니다.",
+      "card": {
+        "session": "sess_24800_1783060135637_1783060135875_396d1f1283f84451dc348604178b23f8",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 150321,
+        "turns": 9,
+        "tool_calls": 10,
+        "tools_per_turn": 1.1,
+        "total_tool_ms": 684,
+        "llm_stream_ms": 149263,
+        "cache_ratio": 0.7854485808851528,
+        "cached_tokens": 214528,
+        "prompt_tokens": 273128,
+        "uncached_tokens": 58600,
+        "output_tokens": 6747,
+        "cache_weighted_input_10": 80052.8,
+        "cache_weighted_input_25": 112232,
+        "prompt_growth": 43499,
+        "antipatterns": 1,
+        "issues": 12,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "service_tier",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-impl-usage-summary#1",
+      "agent": "worker",
+      "ok": true,
+      "ms": 81842,
+      "sessionId": "sess_22404_1783060135650_1783060135889_630e370d8ff15d6c58914f3e6ade30c3",
+      "output": "sessionId: sess_22404_1783060135650_1783060135889_630e370d8ff15d6c58914f3e6ade30c3\n요청 범위가 작고 실행 승인이 포함되어 있어 바로 반영하고 검증하겠습니다.이미 해당 스크립트와 npm 엔트리가 있어 요구사항과 동작을 확인한 뒤 필요한 경우만 보정하겠습니다.필요 구현은 이미 작업트리에 반영된 상태라 검증을 실행하겠습니다.완료했습니다, 재영님.\n\n- **Diff 요약**: `package.json`에 `usage:today` 추가, `scripts/usage-today.mjs` 신규 추가.\n- **기능**: `MIXDOG_DATA_DIR` 우선으로 `history/agent-trace.jsonl`을 읽고, 로컬 자정 이후 `usage` rows를 모델별/총합으로 집계합니다.\n- **검증**: `node --check scripts/usage-today.mjs` 통과, `npm run usage:today` 실행 완료.\n- **참고**: 기존 작업트리에 있던 `src/runtime/channels/lib/*` 수정은 건드리지 않았습니다.",
+      "card": {
+        "session": "sess_22404_1783060135650_1783060135889_630e370d8ff15d6c58914f3e6ade30c3",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 71343,
+        "turns": 13,
+        "tool_calls": 15,
+        "tools_per_turn": 1.2,
+        "total_tool_ms": 2475,
+        "llm_stream_ms": 68564,
+        "cache_ratio": 0.74159765963732,
+        "cached_tokens": 155648,
+        "prompt_tokens": 209882,
+        "uncached_tokens": 54234,
+        "output_tokens": 5536,
+        "cache_weighted_input_10": 69798.8,
+        "cache_weighted_input_25": 93146,
+        "prompt_growth": 14590,
+        "antipatterns": 1,
+        "issues": 15,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "cache_miss",
+          "service_tier",
+          "missed_parallelism",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-explain-compact-engine#1",
+      "agent": "explore",
+      "ok": true,
+      "ms": 87968,
+      "sessionId": "sess_10516_1783060135671_1783060135907_3c455a35c2d88303ec51be56689a9e3c",
+      "output": "sessionId: sess_10516_1783060135671_1783060135907_3c455a35c2d88303ec51be56689a9e3c\n완료했습니다, 재영님. compaction subsystem 요약입니다.\n\n- **진입점:** `engine.mjs`는 `semanticCompactMessages`와 `recallFastTrackCompactMessages`를 내보내고, `compact.mjs` facade가 그대로 re-export합니다. `src/runtime/agent/orchestrator/session/compact/engine.mjs:278`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:445`, `src/runtime/agent/orchestrator/session/compact.mjs:52`\n\n- **주요 호출자:** live loop는 pre-send에서 정책을 계산해 semantic/recall 경로를 호출하고, manager의 manual/auto-clear 경로는 `runSessionCompaction`을 통해 같은 엔진을 호출합니다. `src/runtime/agent/orchestrator/session/loop.mjs:329`, `src/runtime/agent/orchestrator/session/loop.mjs:414`, `src/runtime/agent/orchestrator/session/loop.mjs:441`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:143`\n\n- **발화 조건:** auto가 켜져 있고 boundary가 있으면 trigger는 explicit sub-boundary limit 또는 boundary-buffer로 정해지며, pressure가 trigger 이상이거나 reactive overflow면 compact가 실행됩니다. `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:89`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:107`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:117`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:170`\n\n- **실행 흐름:** loop는 queued steering까지 반영한 뒤 토큰을 추정하고, compact 대상이면 `compacting` stage/telemetry 후 recall-fasttrack 또는 semantic 중 하나를 실행합니다. `src/runtime/agent/orchestrator/session/loop.mjs:324`, `src/runtime/agent/orchestrator/session/loop.mjs:351`, `src/runtime/agent/orchestrator/session/loop.mjs:373`, `src/runtime/agent/orchestrator/session/loop.mjs:414`\n\n- **semantic window:** 먼저 tool-pair sanitize/dedup 후 protected prefix, live, previous summary를 분리하고, user turn이 있으면 최근 user turns 기준, 없으면 tool-result로 시작하지 않는 cut point 기준으로 head/tail을 자릅니다. `src/runtime/agent/orchestrator/session/compact/engine.mjs:100`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:121`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:235`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:240`\n\n- **semantic tail:** 기본 최근 2 turns를 `preserveRecentBudget` 안에서 보존하고, 최신 turn이 너무 크면 suffix만 보존하며, no-user transcript는 valid cut을 최신→과거로 확장하다 초과 직전에 멈춥니다. `src/runtime/agent/orchestrator/session/compact/engine.mjs:46`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:134`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:151`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:193`\n\n- **recall tail:** recall-fasttrack은 최신 user-anchored turns 최대 2개를 구조 그대로 보존하고, cap 초과 시 오래된 turn을 버린 뒤 단일 turn을 truncate하며 no-user tail도 허용합니다. `src/runtime/agent/orchestrator/session/compact/engine.mjs:449`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:593`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:601`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:613`\n\n- **redaction hook:** fits-budget no-op이 아니면 window selection 전에 `redactToolCallSecretsInMessages`가 실행되어 preserved system/tail에 남는 tool-call argument 값만 shape-preserving 방식으로 scrub됩니다. `src/runtime/agent/orchestrator/session/compact/engine.mjs:286`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:306`, `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:219`, `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:323`\n\n- **semantic output:** mandatory(system+tail)가 target보다 크면 budget을 summary room까지 올리고, fitted prompt를 provider에 보내 summary schema를 repair한 뒤 summary+tail 결과가 budget을 넘으면 실패합니다. `src/runtime/agent/orchestrator/session/compact/engine.mjs:319`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:328`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:333`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:371`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:392`\n\n- **recall pipeline:** loop recall-fasttrack은 session memory ingest/dump/cycle1 drain을 시도하고, 실패한 ingest/cycle1은 diagnostics로 남긴 뒤 combined recall text를 `recallFastTrackCompactMessages`에 주입합니다. `src/runtime/agent/orchestrator/session/loop/recall-fasttrack.mjs:61`, `src/runtime/agent/orchestrator/session/loop/recall-fasttrack.mjs:84`, `src/runtime/agent/orchestrator/session/loop/recall-fasttrack.mjs:105`, `src/runtime/agent/orchestrator/session/loop/recall-fasttrack.mjs:148`, `src/runtime/agent/orchestrator/session/loop/recall-fasttrack.mjs:161`\n\n- **fallback:** loop semantic 실패는 non-recall일 때 `pruneToolOutputsUnanchored`로 한 번 salvage하고, salvage 실패 시 overflow telemetry 후 `agentContextOverflowError`를 던집니다. `src/runtime/agent/orchestrator/session/loop.mjs:500`, `src/runtime/agent/orchestrator/session/loop.mjs:517`, `src/runtime/agent/orchestrator/session/loop.mjs:527`, `src/runtime/agent/orchestrator/session/loop.mjs:552`, `src/runtime/agent/orchestrator/session/loop.mjs:618`\n\n- **manager fallback:** manager recall-fasttrack 실패는 semantic fallback을 한 번 시도하고, non-recall 실패는 prune fallback을 시도하며, 그래도 실패하면 `lastStage`를 `manual_failed`/`post_turn_failed`로 기록합니다. `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:231`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:242`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:322`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:348`\n\n- **reactive path:** provider context-overflow refusal은 한 번만 `reactiveOverflowRetryPending`을 세워 다음 loop에서 동일 pre-send compact path를 강제하고, 재실패하면 overflow를 surface합니다. `src/runtime/agent/orchestrator/session/loop.mjs:982`, `src/runtime/agent/orchestrator/session/loop.mjs:995`, `src/runtime/agent/orchestrator/session/loop.mjs:1002`, `src/runtime/agent/orchestrator/session/loop.mjs:1018`",
+      "card": {
+        "session": "sess_10516_1783060135671_1783060135907_3c455a35c2d88303ec51be56689a9e3c",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 77458,
+        "turns": 9,
+        "tool_calls": 12,
+        "tools_per_turn": 1.3,
+        "total_tool_ms": 1160,
+        "llm_stream_ms": 75762,
+        "cache_ratio": 0.6091671542812561,
+        "cached_tokens": 124928,
+        "prompt_tokens": 205080,
+        "uncached_tokens": 80152,
+        "output_tokens": 6295,
+        "cache_weighted_input_10": 92644.8,
+        "cache_weighted_input_25": 111384,
+        "prompt_growth": 60721,
+        "antipatterns": 1,
+        "issues": 17,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "cache_miss",
+          "broad_tool_result",
+          "service_tier",
+          "missed_parallelism",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-debug-forwarding-gap#2",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 126411,
+      "sessionId": "sess_35760_1783060217488_1783060217721_527b673486b2ddb9ace2cd8196d6f95a",
+      "output": "sessionId: sess_35760_1783060217488_1783060217721_527b673486b2ddb9ace2cd8196d6f95a\n재영님, 조사 결과 `bridge-state.json`은 `start()` 초반에 먼저 활성화되고(`src/runtime/channels/index.mjs:1734`), 실제 Discord 연결·소유권·transcript binding·watch는 그 뒤에 따라옵니다.\n\n- **최대 손실: 기존 transcript를 EOF에서 바인딩** — `catchUpFromPersisted`가 같은 persisted path가 아니면 cursor를 `currentSize`로 잡아(`src/runtime/channels/lib/output-forwarder.mjs:117`) activation 이후 binding 전 출력도 backlog로 읽지 않습니다. 자동 복구 없음.\n\n- **경로 미발견: watch/rearm 자체 없음** — activation 시 transcript도 previous path도 못 찾으면 rebind가 실패하고(`src/runtime/channels/index.mjs:515`), 나중에 첫 inbound가 기존 파일을 잡아도 일반 bind는 EOF 시작입니다(`src/runtime/channels/index.mjs:1597`). 자동 복구 없음.\n\n- **파일 미생성: fs.watch 실패 구간** — missing file은 `watch()`가 catch되고 닫힙니다(`src/runtime/channels/lib/output-forwarder.mjs:647`). pending path가 있으면 re-arm poll이 생성 후 `forwardNewText()`로 backlog를 회수합니다(`src/runtime/channels/index.mjs:469`), 단 60초 timeout 이후는 복구 없음(`src/runtime/channels/index.mjs:384`).\n\n- **소유권/연결 전 출력 지연** — owner probe가 `bridgeRuntimeConnected && currentOwnerState().owned`라 pre-connect flush는 읽기 전에 반환합니다(`src/runtime/channels/index.mjs:316`, `src/runtime/channels/lib/output-forwarder.mjs:351`). cursor가 전진하지 않아 post-connect 재시도에서 복구됩니다(`src/runtime/channels/index.mjs:801`).\n\n- **watch 설치 race는 방어됨** — `setContext()` stat 후 `watch()` 설치 사이 쓰인 bytes는 즉시 flush 예약으로 회수됩니다(`src/runtime/channels/lib/output-forwarder.mjs:665`); partial line도 newline 전에는 소비하지 않습니다(`src/runtime/channels/lib/output-forwarder.mjs:172`).\n\nVerdict: pending transcript path가 잡힌 경우는 대체로 지연 후 회수되지만, 기존 transcript를 persisted cursor 없이 EOF로 바인딩하는 activation window는 실제 silent drop 가능성이 있습니다.",
+      "card": {
+        "session": "sess_35760_1783060217488_1783060217721_527b673486b2ddb9ace2cd8196d6f95a",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 115932,
+        "turns": 7,
+        "tool_calls": 8,
+        "tools_per_turn": 1.1,
+        "total_tool_ms": 524,
+        "llm_stream_ms": 111128,
+        "cache_ratio": 0.32630401223640043,
+        "cached_tokens": 69120,
+        "prompt_tokens": 211827,
+        "uncached_tokens": 142707,
+        "output_tokens": 8039,
+        "cache_weighted_input_10": 149619,
+        "cache_weighted_input_25": 159987,
+        "prompt_growth": 35107,
+        "antipatterns": 1,
+        "issues": 11,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "broad_tool_result",
+          "service_tier",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-impl-usage-summary#2",
+      "agent": "worker",
+      "ok": true,
+      "ms": 89481,
+      "sessionId": "sess_34948_1783060223630_1783060223848_a086c7c3dc32e8d9d25fdba3a4b15cf9",
+      "output": "sessionId: sess_34948_1783060223630_1783060223848_a086c7c3dc32e8d9d25fdba3a4b15cf9\n요청 범위대로 스크립트 추가와 검증까지 진행하겠습니다.구현 항목은 이미 작업 트리에 반영되어 있어 검증을 실행하겠습니다.관련 변경만 확인해 diff 요약을 정리하겠습니다.완료됨 — `usage:today` 스크립트와 `scripts/usage-today.mjs`가 작업 트리에 반영되어 있고 검증도 통과했습니다.\n\n- **Diff 요약**: `package.json`에 `usage:today` 추가, 새 스크립트는 `MIXDOG_DATA_DIR`/기본 history 경로를 읽어 오늘 usage를 모델별·총합으로 집계합니다.\n- **검증**: `node --check scripts/usage-today.mjs` 통과, `npm run usage:today` 실행 완료.\n- **참고**: 현재 `src/runtime/channels/lib/event-queue.mjs`, `src/runtime/channels/lib/output-forwarder.mjs`의 기존 수정은 이번 작업에서 건드리지 않았습니다.",
+      "card": {
+        "session": "sess_34948_1783060223630_1783060223848_a086c7c3dc32e8d9d25fdba3a4b15cf9",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 79007,
+        "turns": 11,
+        "tool_calls": 15,
+        "tools_per_turn": 1.4,
+        "total_tool_ms": 2544,
+        "llm_stream_ms": 76106,
+        "cache_ratio": 0.8223639038744482,
+        "cached_tokens": 134144,
+        "prompt_tokens": 163120,
+        "uncached_tokens": 28976,
+        "output_tokens": 6085,
+        "cache_weighted_input_10": 42390.4,
+        "cache_weighted_input_25": 62512,
+        "prompt_growth": 15753,
+        "antipatterns": 1,
+        "issues": 14,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "service_tier",
+          "missed_parallelism",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-explain-compact-engine#2",
+      "agent": "explore",
+      "ok": true,
+      "ms": 73478,
+      "sessionId": "sess_44084_1783060296498_1783060296721_5059cdd1f7776ce3ddcb030a83e15e87",
+      "output": "sessionId: sess_44084_1783060296498_1783060296721_5059cdd1f7776ce3ddcb030a83e15e87\nArchitecture brief, anchored:\n\n- `engine.mjs` exports two compactors: `semanticCompactMessages` and `recallFastTrackCompactMessages`. `src/runtime/agent/orchestrator/session/compact/engine.mjs:278`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:445`\n- Live auto-compaction enters from `agentLoop`, reached by `askSession`; manual/API compaction enters through `runSessionCompaction`. `src/runtime/agent/orchestrator/session/loop.mjs:441`, `src/runtime/agent/orchestrator/session/manager.mjs:1379`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:143`\n- Compaction fires pre-send after queued steering is drained, then only when policy is auto-enabled and pressure reaches trigger, tokens are unknown, or reactive overflow forces it. `src/runtime/agent/orchestrator/session/loop.mjs:324`, `src/runtime/agent/orchestrator/session/loop.mjs:329`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:170`\n- Policy resolves boundary, trigger, reserve, type, semantic/recall flags, and preserved-tail budgets from session/env/tool reserve. `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:89`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:124`\n- Runtime chooses recall-fasttrack first when configured, otherwise semantic; unsupported types fail. `src/runtime/agent/orchestrator/session/loop.mjs:414`, `src/runtime/agent/orchestrator/session/loop.mjs:439`, `src/runtime/agent/orchestrator/session/loop.mjs:496`\n- Semantic compaction sanitizes tool pairs, splits protected prefix/live/previous summary, then redacts tool-call secrets before window selection. `src/runtime/agent/orchestrator/session/compact/engine.mjs:284`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:100`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:312`\n- Semantic preserved tail is chosen by recent user turns if present, otherwise by non-tool cut points; tail is re-sanitized so tool pairs stay valid. `src/runtime/agent/orchestrator/session/compact/engine.mjs:240`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:121`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:250`\n- Semantic output is protected prefix + fitted summary + preserved tail, with mandatory tail/system cost allowed to raise budget. `src/runtime/agent/orchestrator/session/compact/engine.mjs:319`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:328`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:392`\n- Recall-fasttrack preserves the newest structured turns, drops oldest turns over cap, then truncates oldest surviving content if needed. `src/runtime/agent/orchestrator/session/compact/engine.mjs:449`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:593`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:601`\n- Recall-fasttrack also redacts before splitting context, then emits protected prefix + recall summary + preserved tail. `src/runtime/agent/orchestrator/session/compact/engine.mjs:644`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:646`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:695`\n- Redaction is shape-preserving for verbatim retained tool-call args and raw secret strings; prompt-side preserved facts also redact raw secrets. `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:217`, `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:230`, `src/runtime/agent/orchestrator/session/compact/budget.mjs:152`\n- Semantic no-op returns sanitized messages without preserved-tail redaction; redaction only hooks into real compaction. `src/runtime/agent/orchestrator/session/compact/engine.mjs:286`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:306`\n- Failures include empty live/head, missing provider, prompt/summary fit failures, empty provider response, and final budget overflow. `src/runtime/agent/orchestrator/session/compact/engine.mjs:254`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:279`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:335`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:371`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:394`\n- Loop fallback: semantic failures can recover via unanchored tool-output prune; unrecovered failures emit telemetry and raise context overflow. `src/runtime/agent/orchestrator/session/loop.mjs:501`, `src/runtime/agent/orchestrator/session/loop.mjs:517`, `src/runtime/agent/orchestrator/session/loop.mjs:552`, `src/runtime/agent/orchestrator/session/loop.mjs:618`\n- Manual/session fallback: recall-fasttrack failure may degrade to semantic; non-recall failures may prune; otherwise result records error without replacing history. `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:237`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:245`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:322`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:348`",
+      "card": {
+        "session": "sess_44084_1783060296498_1783060296721_5059cdd1f7776ce3ddcb030a83e15e87",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 62943,
+        "turns": 8,
+        "tool_calls": 8,
+        "tools_per_turn": 1,
+        "total_tool_ms": 735,
+        "llm_stream_ms": 61541,
+        "cache_ratio": 0.6553850865104884,
+        "cached_tokens": 107008,
+        "prompt_tokens": 163275,
+        "uncached_tokens": 56267,
+        "output_tokens": 5391,
+        "cache_weighted_input_10": 66967.8,
+        "cache_weighted_input_25": 83019,
+        "prompt_growth": 38835,
+        "antipatterns": 1,
+        "issues": 14,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "cache_miss",
+          "broad_tool_result",
+          "service_tier",
+          "missed_parallelism",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    }
+  ],
+  "task_complete": true,
+  "task_errors": [],
+  "score_complete": true,
+  "score_errors": [],
+  "group": {
+    "n": 6,
+    "wall_ms": 92834,
+    "turns": 9.5,
+    "tool_calls": 11.3,
+    "tools_per_turn": 1.2,
+    "total_tool_ms": 1353.7,
+    "llm_stream_ms": 90394,
+    "cache_ratio": 0.7,
+    "cached_tokens": 134229.3,
+    "prompt_tokens": 204385.3,
+    "uncached_tokens": 70156,
+    "output_tokens": 6348.8,
+    "cache_weighted_input_10": 83578.9,
+    "cache_weighted_input_25": 103713.3,
+    "prompt_growth": 34750.8,
+    "antipatterns": 1,
+    "issues": 13.8,
+    "grep_sweep_count": 0
+  },
+  "cards": [
+    {
+      "session": "sess_24800_1783060135637_1783060135875_396d1f1283f84451dc348604178b23f8",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 150321,
+      "turns": 9,
+      "tool_calls": 10,
+      "tools_per_turn": 1.1,
+      "total_tool_ms": 684,
+      "llm_stream_ms": 149263,
+      "cache_ratio": 0.7854485808851528,
+      "cached_tokens": 214528,
+      "prompt_tokens": 273128,
+      "uncached_tokens": 58600,
+      "output_tokens": 6747,
+      "cache_weighted_input_10": 80052.8,
+      "cache_weighted_input_25": 112232,
+      "prompt_growth": 43499,
+      "antipatterns": 1,
+      "issues": 12,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "service_tier",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_22404_1783060135650_1783060135889_630e370d8ff15d6c58914f3e6ade30c3",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 71343,
+      "turns": 13,
+      "tool_calls": 15,
+      "tools_per_turn": 1.2,
+      "total_tool_ms": 2475,
+      "llm_stream_ms": 68564,
+      "cache_ratio": 0.74159765963732,
+      "cached_tokens": 155648,
+      "prompt_tokens": 209882,
+      "uncached_tokens": 54234,
+      "output_tokens": 5536,
+      "cache_weighted_input_10": 69798.8,
+      "cache_weighted_input_25": 93146,
+      "prompt_growth": 14590,
+      "antipatterns": 1,
+      "issues": 15,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "cache_miss",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_10516_1783060135671_1783060135907_3c455a35c2d88303ec51be56689a9e3c",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 77458,
+      "turns": 9,
+      "tool_calls": 12,
+      "tools_per_turn": 1.3,
+      "total_tool_ms": 1160,
+      "llm_stream_ms": 75762,
+      "cache_ratio": 0.6091671542812561,
+      "cached_tokens": 124928,
+      "prompt_tokens": 205080,
+      "uncached_tokens": 80152,
+      "output_tokens": 6295,
+      "cache_weighted_input_10": 92644.8,
+      "cache_weighted_input_25": 111384,
+      "prompt_growth": 60721,
+      "antipatterns": 1,
+      "issues": 17,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "cache_miss",
+        "broad_tool_result",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_35760_1783060217488_1783060217721_527b673486b2ddb9ace2cd8196d6f95a",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 115932,
+      "turns": 7,
+      "tool_calls": 8,
+      "tools_per_turn": 1.1,
+      "total_tool_ms": 524,
+      "llm_stream_ms": 111128,
+      "cache_ratio": 0.32630401223640043,
+      "cached_tokens": 69120,
+      "prompt_tokens": 211827,
+      "uncached_tokens": 142707,
+      "output_tokens": 8039,
+      "cache_weighted_input_10": 149619,
+      "cache_weighted_input_25": 159987,
+      "prompt_growth": 35107,
+      "antipatterns": 1,
+      "issues": 11,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "broad_tool_result",
+        "service_tier",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_34948_1783060223630_1783060223848_a086c7c3dc32e8d9d25fdba3a4b15cf9",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 79007,
+      "turns": 11,
+      "tool_calls": 15,
+      "tools_per_turn": 1.4,
+      "total_tool_ms": 2544,
+      "llm_stream_ms": 76106,
+      "cache_ratio": 0.8223639038744482,
+      "cached_tokens": 134144,
+      "prompt_tokens": 163120,
+      "uncached_tokens": 28976,
+      "output_tokens": 6085,
+      "cache_weighted_input_10": 42390.4,
+      "cache_weighted_input_25": 62512,
+      "prompt_growth": 15753,
+      "antipatterns": 1,
+      "issues": 14,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_44084_1783060296498_1783060296721_5059cdd1f7776ce3ddcb030a83e15e87",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 62943,
+      "turns": 8,
+      "tool_calls": 8,
+      "tools_per_turn": 1,
+      "total_tool_ms": 735,
+      "llm_stream_ms": 61541,
+      "cache_ratio": 0.6553850865104884,
+      "cached_tokens": 107008,
+      "prompt_tokens": 163275,
+      "uncached_tokens": 56267,
+      "output_tokens": 5391,
+      "cache_weighted_input_10": 66967.8,
+      "cache_weighted_input_25": 83019,
+      "prompt_growth": 38835,
+      "antipatterns": 1,
+      "issues": 14,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "cache_miss",
+        "broad_tool_result",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    }
+  ]
+}